# loss function
lossDict = {}
# objective function => 이것을 이용해서 loss를 구하거나 함.
objDict={}
# X0, X1: Gray frames
# Y0, Y1: Ground truth color frames
# C0, C1: Colorized frames
with tf.compat.v1.variable_scope(tf.compat.v1.get_variable_scope()):
X0, X1 = input_i[:,:,:,0:1], input_i[:,:,:,1:2]
Y0, Y1 = input_target[:,:,:,0:3], input_target[:,:,:,3:6]
# colorization network 구축
with tf.compat.v1.variable_scope('individual'):
C0=net.VCN(utils.build(tf.tile(X0, [1,1,1,3])),reuse=False, div_num=4)
C1=net.VCN(utils.build(tf.tile(X1, [1,1,1,3])),reuse=True, div_num=4)
objDict["mask"],_=occlusion_mask(Y0,Y1,input_flow_backward[:,:,:,0:2])
# objDict["warped"]=flow_warp_op(C0,input_flow_backward[:,:,:,0:2])
# objDict["warped"] = tfa.image.dense_image_warp(C0, input_flow_backward[:,:,:,0:2], name="op_warp")
# objDict에 warped를 저장할 때, C0를 warp 시키는 것이기 때문에 C0의 output channel에 맞춰서 shape를 맞춰줘야 함.
# 혹시 dense_image_warp의 channel이 3이어야만 한다면, 학습에 영향을 주는 것 같다면 for문을 div_num만큼 돌려서 concat(axis=3)하고 set_shape하기
temp_warped = tfa.image.dense_image_warp(C0, input_flow_backward[:,:,:,0:2], name="op_warp")
temp_warped.set_shape([None, None, None, 3*div_num])
objDict["warped"]= temp_warped
lossDict["RankDiv_im1"]=loss.RankDiverse_loss(C0, tf.tile(input_target[:,:,:,0:3], [1,1,1,div_num]),div_num)
lossDict["RankDiv_im2"]=loss.RankDiverse_loss(C1, tf.tile(input_target[:,:,:,3:6], [1,1,1,div_num]),div_num)
lossDict["RankDiv"]=lossDict["RankDiv_im1"]+lossDict["RankDiv_im2"]
# loss function
lossDict = {}
# objective function => 이것을 이용해서 loss를 구하거나 함.
objDict = {}
# X0, X1: Gray frames
# Y0, Y1: Ground truth color frames
# C0, C1: Colorized frames
with tf.compat.v1.variable_scope(tf.compat.v1.get_variable_scope()):
X0, X1 = input_i[:, :, :, 0:1], input_i[:, :, :, 1:2]
Y0, Y1 = input_target[:, :, :, 0:3], input_target[:, :, :, 3:6]
# colorization network 구축
with tf.compat.v1.variable_scope('individual'):
C0 = net.VCN(utils.build(tf.tile(X0, [1, 1, 1, 3])),
reuse=False,
div_num=4)
C1 = net.VCN(utils.build(tf.tile(X1, [1, 1, 1, 3])),
reuse=True,
div_num=4)
objDict["mask"], _ = occlusion_mask(Y0, Y1, input_flow_backward[:, :, :,
0:2])
# objDict["warped"]=flow_warp_op(C0,input_flow_backward[:,:,:,0:2])
# objDict["warped"] = tfa.image.dense_image_warp(C0, input_flow_backward[:,:,:,0:2], name="op_warp")
# objDict에 warped를 저장할 때, C0를 warp 시키는 것이기 때문에 C0의 output channel에 맞춰서 shape를 맞춰줘야 함.
# 혹시 dense_image_warp의 channel이 3이어야만 한다면, 학습에 영향을 주는 것 같다면 for문을 div_num만큼 돌려서 concat(axis=3)하고 set_shape하기
temp_warped = tfa.image.dense_image_warp(C0,
input_flow_backward[:, :, :, 0:2],
activation=None,
kernel_initializer=tf.compat.v1.keras.initializers.VarianceScaling(
scale=1.0, mode="fan_avg", distribution="uniform"),
name=ext + 'g_conv9_2')
return conv9
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.compat.v1.Session(config=config)
input_i = tf.compat.v1.placeholder(tf.float32, shape=[1, None, None, 2])
input_target = tf.compat.v1.placeholder(tf.float32, shape=[1, None, None, 6])
with tf.compat.v1.variable_scope(tf.compat.v1.get_variable_scope()):
with tf.compat.v1.variable_scope('individual'):
g0 = VCN(utils.build(tf.tile(input_i[:, :, :, 0:1], [1, 1, 1, 3])),
reuse=False)
g1 = VCN(utils.build(tf.tile(input_i[:, :, :, 1:2], [1, 1, 1, 3])),
reuse=True)
saver = tf.compat.v1.train.Saver(max_to_keep=1000)
# +
sess.run([tf.compat.v1.global_variables_initializer()])
var_restore = [v for v in tf.compat.v1.trainable_variables()]
saver_restore = tf.compat.v1.train.Saver(var_restore)
ckpt = tf.train.get_checkpoint_state(model)
saver_restore.restore(sess, ckpt.model_checkpoint_path)
# -